Survey of Bias Mitigation in Federated Learning

Lynda Ferraguig (1) , Yasmine Djebrouni (2, 3) , Sara Bouchenak (4, 5, 6) , Vania Marangozova (2, 3)

1 ESI - École Nationale Supérieure d'Informatique [Alger]
2 UGA - Université Grenoble Alpes
3 LIG - Laboratoire d'Informatique de Grenoble
4 INSA Lyon - Institut National des Sciences Appliquées de Lyon
5 LIRIS - Laboratoire d'InfoRmatique en Image et Systèmes d'information
6 DRIM - Distribution, Recherche d'Information et Mobilité

Federated Learning (FL) interestingly allows a set of participants to collectively resolve a machine learning problem in a decentralized and privacy preserving manner. However, data distribution and heterogeneity, that are inherent to FL, may induce and exacerbate the problem of bias, with its prejudicial consequences such as racial or sexist segregation, illegal actions, or reduced revenues. In this paper, we describe the problem of bias in Federated Learning, and provide a comparative review of existing approaches of FL bias mitigation, before discussing open challenges and interesting research directions.

Mots clés: Bias; Federated Learning; Machine Learning

Domaines: Intelligence artificielle [cs.AI]; Calcul parallèle, distribué et partagé [cs.DC]

Lynda Ferraguig, Yasmine Djebrouni, Sara Bouchenak, Vania Marangozova. Survey of Bias Mitigation in Federated Learning. Conférence francophone d'informatique en Parallélisme, Architecture et Système, Jul 2021, Lyon (virtuel), France. ⟨hal-03343288⟩